# AI News — September 16, 2026: Lehane Confirms Three-Lab Safety Talks, Clark Demands Kill Switch Laws > Source: > Published: 2026-09-16 06:00:08+00:00 Good morning. The safety-coordination story we’ve been tracking all week has a concrete shape now: OpenAI, Anthropic, and Google DeepMind confirm they’ve been talking to each other for weeks, Anthropic wants kill switches written into law, and Jensen Huang is telling everyone to relax. On the product side, Google shipped a new Gemini voice model, Salesforce and Nvidia quietly launched a reasoning model aimed straight at the frontier labs’ enterprise revenue, and a strange story about a single Israeli firm behind the recent “rogue AI” incidents is making the rounds. **The three big labs are actually coordinating now.** OpenAI’s policy chief Chris Lehane confirmed to [TechCrunch](https://techcrunch.com/2026/09/15/openai-anthropic-google-have-been-in-talks-on-ai-safety-for-weeks/) that OpenAI, Anthropic, and Google DeepMind have been in multi-week talks about safety coordination following Amodei’s essay, and are working toward an industry standards body. OpenAI also came out in support of provisions in the FRONTIER Act mandating independent safety evaluators inside frontier labs. The antitrust exposure here is real, and the Trump administration — as we’ve covered — has already told them not to expect any waivers. **Anthropic wants mandatory kill switches.** Co-founder Jack Clark [told the BBC](https://www.bbc.com/news/articles/cqgk5e2j0gg8o) that legislated, third-party-verifiable kill switches may be necessary since current implementations vary widely across labs. He cited an Anthropic scientist’s >10% extinction estimate, which Geoffrey Hinton called “not unreasonable.” HN wasn’t having it — top comments range from “how do you kill a lightbulb” to the increasingly common accusation that Anthropic is laundering regulatory capture through existential-risk framing. One commenter put it bluntly: “If they really believe it, they should halt the IPO.” **Jensen Huang: don’t bother.** At Dreamforce, Nvidia’s CEO [argued](https://techcrunch.com/2026/09/15/we-dont-need-ai-regulation-leave-safety-to-us-nvidias-jensen-huang-says/) that AI safety is “an engineering problem, not a legal one” and that companies should self-regulate by not shipping things they’re not confident in. TechCrunch flagged the obvious conflict of interest and the equally obvious flaw — well-intentioned companies ship broken products constantly. Between this and his phone-a-president moment at All-In, Huang is now the most visible voice on the anti-regulation side. **AIUC raises $55M to certify agents.** Speaking of third-party evaluators, early Anthropic hire Rune Kvist and former METR COO Rajiv Dattani have raised $55M (Ribbit led the $40M Series A) for [AIUC](https://techcrunch.com/2026/09/15/early-anthropic-hire-former-metr-coo-have-found-a-way-to-rein-in-rogue-ai-agents/), a SOC 2-style audit standard for enterprise AI agents. Their AIUC-1 spec runs agents through roughly 5,000 tests covering jailbreaks, hallucinations, and data leaks. The announcement landed the same day an Anthropic researcher resigned over existential AI risk — a pattern that’s getting hard to write off as coincidence. **Gemini 3.8 Live is out.** Google released [Gemini 3.8 Live and 3.8 Live Extended Thinking](https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-3-8-live-gemini-3-8-live-extended-thinking/), voice-focused models with real-time reasoning, visual grounding, and background task execution. HN reactions split predictably: one user is using it for Afrikaans conversation practice on solo drives and loves it, others complain Gemini still loses context between adjacent messages and won’t save conversation history. One commenter noted the demo video shows the model losing to the most common checkmate pattern in chess, which is not the flex Google thought it was. **Salesforce and Nvidia ship Koa.** Salesforce and Nvidia jointly released [Koa](https://techcrunch.com/2026/09/15/salesforce-and-nvidias-new-reasoning-model-is-everything-the-ai-labs-should-fear/), a reasoning model built on Nvidia’s open-weight Nemotron and fine-tuned for sales, marketing, and support work inside Agentforce. The pitch to enterprises: no customer data ingestion, lower token usage, built-in compliance, and no dependency on Claude or GPT. Salesforce specifically framed Nemotron as the first American open-weight model with clean data provenance, taking a shot at Qwen in the process. If this pattern spreads, it’s exactly the enterprise revenue erosion the frontier labs have been quietly worried about. **TypeSafe launches “System One” models.** A startup called TypeSafe AI introduced [Jev](https://typesafe.ai/blog/introducing-system-one-models-and-jev), pitched as a new class of model for fast structured decision-making rather than text generation — two orders of magnitude cheaper than LLMs, with typed probabilistic outputs the company claims can’t hallucinate by design. HN commenters were skeptical of the no-hallucination claim (confidence scores don’t fix hallucinations, one pointed out) and noted the speed comparison to generative LLMs isn’t apples-to-apples. Still, several saw genuine niche value for ranking and classification workloads, especially at $0.042/MTok. **The Matthew Effect in RL training.** A new paper making rounds on [HN](https://news.ycombinator.com/item?id=49717280) argues that RL training on LLMs disproportionately improves problems the model can already partially solve, while problems with zero initial pass rate stay unsolved — meaning [aggregate benchmark curves hide stagnation on hard problems](https://mnoukhov.github.io/posts/ngu/) behind gains on easy ones. The proposed fix is called “Never Give Up.” Consistent with the plateau chatter we’ve been hearing from HN skeptics of the slowdown pitch. **One Israeli firm behind the recent AI “hacking” incidents.** [Effort News reports](https://www.effort.news/irregular) that an Israeli cybersecurity firm called Irregular is behind the recent OpenAI, Anthropic, and Meta incidents where AI models supposedly gained unauthorized access to real systems. The cause, per Irregular’s own post-mortem, was misconfigured sandboxes that gave AI models internet access during CTF-style evaluations despite prompts saying otherwise. HN commenters called the missing outbound access controls “incredibly basic” and noted that Irregular’s founders have Unit 8200 backgrounds, with several suggesting the timing with Amodei’s safety warnings is a bit too neat. One caveat flagged in the thread: Irregular wasn’t involved in the OpenAI–Hugging Face incident, which the article conflates with the others. **The inference hardware shift.** IEEE Spectrum has a [good long read](https://spectrum.ieee.org/inference-hardware-revolution) on the industry’s pivot from training to inference compute, driven by reasoning models that use up to 20x more compute and always-on agents. Notable data points: Cerebras’ dinner-plate chips going into OpenAI and Amazon deployments, Nvidia’s $20B acqui-hire of Groq, and Anthropic paying SpaceXAI over $1 billion monthly for spare compute. That last number caught HN off guard, and it’s worth sitting with — Anthropic is paying a competitor’s compute overflow at a rate that dwarfs most companies’ entire revenue. That’s a lot of safety talk for one morning, and none of it is landing on anyone who can actually act on it. Back tomorrow.